Machine learning on asynchronous clinical pages to predict clinical deterioration

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Patient safety

Background:

• Traditional early warning scores for clinical deterioration are limited by delayed or inconsistent documentation.
• Real-time clinical data can potentially improve the prediction of patient deterioration.

Purpose of the Study:

• To develop and evaluate a machine learning pipeline for predicting clinical deterioration using real-time clinician pager messages.
• To compare the performance of this pipeline against existing methods like the Epic Deterioration Index.

Main Methods:

• A retrospective study analyzed adult non-ICU hospitalizations, extracting content and frequency from pager messages.
• Long short-term memory (LSTM) models were trained to predict deterioration events (rapid response, unplanned ICU transfer, cardiac arrest) within 6, 12, or 24 hours.
• Model performance was compared to the Epic Deterioration Index (EDI) and a combined logistic regression ensemble model.

Main Results:

• The LSTM model achieved moderate discriminative ability (AUROC 0.684–0.724) for predicting deterioration within 24 hours.
• Pager message predictions showed moderate correlation with EDI scores, suggesting complementary information.
• The ensemble model integrating pager data and EDI scores consistently outperformed individual methods, achieving an AUROC of 0.797 for 12-hour prediction.

Conclusions:

• Clinical pager messages are an underutilized data source reflecting real-time clinician observations.
• Machine learning applied to pager messages can augment early warning systems, providing timely and intuitive alerts.
• This approach offers a novel way to enhance patient safety by leveraging existing communication data.